PAN at CLEF 2018

In this talk I will review some recent results regarding early detection of signs of depression and anorexia. Since 2017, we have been organizing eRisk, a CLEF lab that promotes the development of effective and efficient solutions for early risk prediction on the Internet. eRisk explores the evaluation methodology, effectiveness metrics and practical applications (particularly those related to health and safety) of early risk detection on the Internet. Early detection technologies can be employed in different areas, particularly those related to health and safety. For instance, early alerts could be sent when a predator starts interacting with a child for sexual purposes, or when a potential offender starts publishing antisocial threats on a blog, forum or social network. Our main goal is to pioneer a new interdisciplinary research area that would be potentially applicable to a wide variety of situations and to many different personal profiles. Examples include potential paedophiles, stalkers, individuals that could fall into the hands of criminal organisations, people with suicidal inclinations, or people susceptible to depression. In this talk, I will discuss the lessons learned over these two years and some future lines of work.

The drastic change in the Web was witnessed throughout the past decade, which saw an exponential growth in social networking services. Traditionally, social network users are encouraged to complete their profiles by explicitly providing their personal attributes such as age, gender, interests, etc. Such information is essential for Marketing, Facility Arrangement, or Candidate Assessment, but, unfortunately, often not publicly available. This gives rise to user profiling, which aims at automatic inference of individual user attributes based on their social network interactions. Considering that human beings frequently contribute multi-modal data in multiple online social networks at the same time, it is essential to implement inter-source complimentary multi-view learning techniques to perform automatic user profiling efficiently. In this talk, we will overview recent research attempts on learning across multiple social networks and data modalities for automatic user profiling. We will also give several practical examples of how Multi-View User Profiling helps SoMin.ai in boosting the efficiency of enterprises' marketing efforts.

There is much concern about algorithms that underlie information services and the view of the social world they present to users. Image search engines are known to perpetuate gender stereotypes, particularly surrounding professions (e.g., returning primarily images of men on a search for "engineer," although few, if any, men on a search for "nurse"). In the first part of the talk, I discuss the problem of detecting social biases in image search results. We developed a novel method for automatically examining the content and strength of gender stereotypes in image results, which is inspired by the trait adjective checklist method. In experiments with Microsoft Bing, we found that photos of women are more often retrieved for searches on warm character traits (e.g., "emotional"), whereas agentic traits (e.g., "rational") typically result in more images of men. In the second part of the talk, I address questions surrounding the origin of social biases in search algorithms. I will argue that the quality of image metadata is a source of bias, as algorithms are typically trained on "gold standard," human-produced metadata. Specifically, in an experiment testing a commonly used crowdsourcing task for metadata generation, I will provide evidence that people's descriptions of men and women depicted in similar contexts differ in systematic ways that are predictable by theory. In conclusion, I shall argue that while the reproduction of social stereotypes in search algorithms is likely inevitable, there are ways to effectively raise users' awareness of biases in results.